Impact of Data Loss in Postprocessing on Training and Inference of Quantum Neural Networks

In the post-processing stage of quantum neural networks, data loss due to invalid assumptions significantly affects training and inference performance. Specifically, the \tt{SamplerQNN} class in the Qiskit library causes 85% to 99.6% of valid measurement data to be lost when handling IBM hardware with more than 100 physical quantum bits, due to its filtering method based on the virtual quantum bit space. This defect results in unnormalized probability vectors, leading to prediction distortions and incorrect propagation of loss values, and the API fails to issue warnings. In five experimental tests, the inference accuracy dropped sharply from 0.94 to 0.39; during training, the loss signal was compressed by 22 to 27 times, significantly reducing the optimizer’s sensitivity to the target landscape. This issue exists in all released versions of the library from version 0.8.4 to 0.9. Currently, the community has implemented a layout-based edge-case repair solution (GitHub Pull Request #1041) to address the \tt{Sampl…